Collaborate on a frontier RTL handoff study
SV-Gap is inviting frontier-model and chip-design AI researchers to co-design a shared evaluation layer for the questions that functional benchmarks leave unanswered. Participation does not require publishing a private checkpoint name, provider transcript, internal harness, or proprietary RTL.
The smallest useful starting point
Start with one 30-minute research scoping call. The public intake contains only a non-confidential question; scheduling is coordinated separately. The call is not a request for code and creates no obligation to publish or continue.
| Path | What you provide | What the project returns |
|---|---|---|
| Research question | One production question that functional evaluation does not answer | A falsifiable experiment, claim boundary, and smallest publishable evidence unit |
| Model study | One model or agent surface that could run through a public task | A smoke/full protocol choice and a portable evidence-profile plan |
| Design partner | One applied handoff problem that can be represented with public or synthetic artifacts | A bounded workflow from gap map to reproducible result |
| Method co-design | A task, checker result, benchmark surface, or disputed finding | A scoped RFC, reproducibility boundary, and contribution path |
Start with the structured collaboration intake. It requires no installation and is the preferred path for new collaborators. If GitHub sign-in is a blocker, email the maintainer with one public, non-confidential sentence about the question. Applied teams can review the complete design-partner workflow before deciding whether to continue after the call.
Ways to participate
Run the frozen generation taskpack
Connect an existing model harness through the stdin/stdout command contract, run reset-release v0.2, and report functional, structural, unknown, and tool error outcomes. Stable attested aliases are supported.
Contribute a diagnosis or repair replication
Use the public starter workflows to test whether an agent preserves epistemic uncertainty or removes a finding without functional regression.
Co-design the evidence contract
Propose an intent-bearing digital RTL taskpack, independent open checker, benchmark adapter, disputed finding, or evidence field needed by a downstream production review.
What collaborators receive
- a stable, citable evidence-profile URL for accepted submissions;
- visible repository and artifact credit;
- explicit claim and provenance boundaries;
- review of schema, reproducibility, and public-artifact safety; and
- the opportunity to shape the next frozen protocol through an RFC.
An intake that is outside the current scope will still receive a direct fit assessment and the missing capability or evidence, rather than a generic invitation to contribute.
Research-paper collaboration is possible but not promised by participation. Authorship is decided separately based on substantive intellectual contribution, analysis, drafting, and accountability under the venue's policy. Joining a call or providing an affiliation does not create contributor status. Named contributor credit requires an attributable accepted contribution such as protocol design, a redistributable result, task design, analysis, validation, documentation, or code.
Response expectations
The maintainer will acknowledge a scoped research question, result-submission proposal, or reproducible oracle dispute within two working days. Acknowledging a proposal is not the same as accepting its claim. If review will take longer, the thread will state the missing evidence and next decision point.
Start with the collaboration intake. The maintainer email is the no-GitHub-account alternative. Use the replication and co-design discussion for an open-ended research thread after the initial scope is clear. Do not post proprietary RTL, credentials, confidential constraints, private checkpoint mappings, or unapproved tool output.